<p>Concrete crack classification and detection are essential to provide safety and durability of civil structures. Manual inspection techniques are inefficient, error-ridden, and time-consuming, making automated, precise, and scalable solutions inevitable. But complexities in the crack patterns, changing illumination conditions, surface textures, and the availability of noise pose significant challenges for image-based crack detection and classification systems. To solve these problems, introduce a new deep learning framework integrating state-of-the-art segmentation, feature extraction, and detection and classification methods. A DFA-UNet model is employed for accurate crack segmentation using dual feature attention. This improves the localization of crack boundaries and reduces noise disturbances. Then, Principal Component Analysis is applied for dimensionality reduction and best feature extraction to provide computational efficiency while preserving informative data. For detection and classification, construct an Enhanced Cascade Convolutional Neural Network (ECCNN) based on an increase over a cascaded structure for better recognition accuracy. ECCNN performance is further improved by a Cascade CNN with hyperparameter optimization by the Adaptive Osprey Optimization Algorithm for strong convergence and generalization. The experimental result on benchmark datasets demonstrates improved accuracy, efficiency, and robustness of our approach in concrete crack detection and classification, which shows its potential for implementation in real-world structural health monitoring and maintenance.</p>